← AI Terminology
Verifier Model
A verifier model is a model (or program) trained or designed to check whether a candidate answer or reasoning trace is correct, rather than to generate answers from scratch.
Verifiers support search, best-of-N, and safe deployment filters.
Verifiers support search, best-of-N, and safe deployment filters.
Why It Matters in AI
Generators propose; verifiers dispose. Separating creation from checking improves math, code, and factual pipelines and is central to scalable oversight ideas. Many reasoning systems pair a strong generator with a specialised verifier.
Key Points
| Aspect | Description |
|---|---|
| Use | Filter/rank candidates; provide RL rewards |
| Risk | False accepts/rejects; adversarial fooling |
| Forms | ORM/PRM, unit tests, formal checkers, LLM judges |
| Benefit | Can be smaller/specialised yet still improve a larger generator |
| Related | Process supervision, debate, recursive reward modeling |
| Oversight | Humans may only audit verifier disagreements |
Simple Analogy
A proofreader who does not write the book but catches errors — the generator authors, the verifier stamps approve/reject.
Common Usage Examples
- Math verifier checks final numeric answer
- Code tests as programmatic verifiers
- LLM-as-judge pairwise comparison
- Search: expand only steps the PRM verifier likes
Summary
In short: A verifier model checks candidates instead of writing them — the ranking and filtering engine behind many high-reliability AI systems.